| ### | |
| # | |
| # Slurm resctrl groups configuration file for the task/resctrl plugin | |
| # | |
| # See man resctrl.yaml and https://slurm.schedmd.com/resctrl.html for | |
| # further information. | |
| # | |
| # The three groups below must all be defined. Every step of a job is placed | |
| # in one of them according to the GPUs allocated to the job on the node: | |
| # none goes to cpu_only, at most one on any socket to gpu_one_per_socket, | |
| # and two or more on one socket to gpu_two_per_socket. Jobs that own the | |
| # whole node are not placed in any group. | |
| # | |
| # The schemata is written to /sys/fs/resctrl/<group>/schemata as is, so the | |
| # values follow the conventions of the hardware. On Intel and Arm MPAM the | |
| # memory bandwidth (MB) is a percentage per memory domain; on AMD it is in | |
| # units of 1/8 GB/s. A value limits the group, not each job, except that | |
| # Intel without mba_MBps throttles each core. This example is for nodes whose | |
| # /sys/fs/resctrl/schemata lists the MB domains 0 and 1. | |
| #-- | |
| groups: | |
| # Jobs without GPUs | |
| - name: cpu_only | |
| schemata: | | |
| MB:0=30;1=30 | |
| # Jobs with at most one GPU on a socket | |
| - name: gpu_one_per_socket | |
| schemata: | | |
| MB:0=35;1=35 | |
| # Jobs with two or more GPUs on a socket | |
| - name: gpu_two_per_socket | |
| schemata: | | |
| MB:0=70;1=70 |